About & methodology

EdgeLLM Bench collects publicly posted local-LLM benchmark results for edge boards. We only record numbers present in the source; fields a source didn't report are simply left out rather than guessed. "Prompt t/s" is prompt processing (usually llama-bench pp512); "Gen t/s" is token generation (usually tg128). Results from different runtimes, threads, and cooling are not strictly comparable. Check each source.

Last verified 2026-10-17: every result was re-checked against its cited source and corrected where it differed; India prices were re-checked on retailer pages the same day (Robu.in blocks automated checks, so its prices are from 2026-10-16).

Affiliate disclosure (placeholder): some "where to buy" links will be affiliate links; they never affect rankings.

Sponsorship: any paid placement is labelled (for example a "Featured" badge on a store row) and never changes benchmark numbers, rankings or price order. See Advertise.

How we label and check results

Every row carries a source badge:

• Community posted by an individual: GitHub repos and gists, Reddit, forums, personal blogs. Self-reported; we check the number matches the source but have not re-run it.
• Vendor published by a board or platform maker (for example NVIDIA Jetson AI Lab, Turing Pi, PicoCluster). Often tuned, may be best-case.
• Reproduced re-run by EdgeLLM Bench on our own hardware. None yet.

⚠ unusual, check source: a row is marked when its generation speed is more than 2x or less than half the median of at least 3 results on the same board and model size class (1–5B, 5–11B, 12B+), compared within the same quant-bit group (≤3, 4–5, 6–8 bit) when there are enough results. Flagged rows are not hidden or removed; they may reflect a different runtime, NPU/GPU use or settings. Hover the marker to see the comparison.

New submissions must include pasted raw output (llama-bench or ollama --verbose); results without it are not added.